Revolutionizing upper gastrointestinal disease diagnosis: The transformative role of artificial intelligence in endoscopy

Abstract

With the rapid advancement of technology, artificial intelligence (AI) has emerged as a transformative force in gastroenterology, particularly in diagnosing upper gastrointestinal diseases such as Barrett's esophagus (BE), esophageal cancer, gastroesophageal reflux disease (GERD), and esophagogastric varices. AI's capabilities in image analysis, classification, detection, and segmentation have significantly improved diagnostic accuracy and efficiency. For BE, AI models achieve high sensitivity and specificity in detecting early neoplastic changes and guiding targeted biopsies. In esophageal cancer, AI enhances early lesion detection, improving intervention success rates. For GERD, AI classifies disease severity based on the Los Angeles grading system and accurately segments lesions. Additionally, AI detects esophagogastric varices and predicts bleeding risks more effectively than traditional methods. Despite these advancements, challenges remain, including the need for high-quality data, multi-center validation, and ensuring AI model interpretability. Future research should address these issues and further integrate AI into clinical practice to optimize patient outcomes. This review highlights AI's transformative impact on upper gastrointestinal disease diagnosis, emphasizing its potential to revolutionize endoscopic practice and improve patient care.

Key Words: Artificial intelligence; Upper gastrointestinal diseases; Endoscopic diagnosis; Deep learning; Esophageal lesions; Gastroesophageal reflux disease

Core Tip: This review highlights the application progress of artificial intelligence in the diagnosis of upper gastrointestinal diseases, emphasizing its potential in improving the accuracy of endoscopic image analysis, early lesion detection, and disease grading. It also addresses the challenges faced in clinical implementation, including data quality, multi-center validation, and model interpretability.

INTRODUCTION

With continual advances in global science and technology, the applications of artificial intelligence (AI) in the medical field have gradually widened, especially in the field of digestive endoscopy, where AI has achieved new breakthroughs and garnered widespread clinical attention[1]. AI technology, with its powerful image-recognition and data-analysis capabilities, has brought about unprecedented changes in the diagnosis and treatment of upper gastrointestinal diseases by using digestive endoscopy. Thus far, the application of AI in endoscopic image analysis has mainly focused on classification, detection, and segmentation[2-4]. Through deep learning (DL) algorithms, AI can quickly and accurately classify lesions in endoscopic images, distinguishing between normal and pathological tissues. AI can also identify early cancerous lesions in real-time, significantly improving the detection rate of lesions. Finally, image-segmentation research has shown that AI can precisely delineate the boundaries of lesion areas, providing important references for subsequent diagnosis and treatment[2-4].

The diagnosis and treatment of upper gastrointestinal diseases heavily rely on digestive endoscopy. However, conventional endoscopic examinations have certain limitations, such as the dependence on the experience and subjective judgment of endoscopists, which can lead to missed diagnoses or misdiagnoses[5]. The emergence of AI-assisted digestive endoscopy provides new approaches and methods to address these limitations. AI systems can analyze endoscopic images in real-time, and assist endoscopists in accurately identifying lesions, thereby improving diagnostic accuracy and efficiency, especially in the detection of early lesions[6]. The functions of AI-assisted endoscopy are illustrated in Figure 1. Unsurprisingly, the role of AI-assisted digestive endoscopy in the diagnosis and treatment of upper gastrointestinal diseases has become a research hot spot.

Figure 1  The functions of artificial intelligence-assisted endoscopy.

In recent years, considerable progress has been made in the research on the use of AI in Barrett’s esophagus (BE); by analyzing subtle features in endoscopic images, AI can accurately identify BE lesions and determine their potential for malignant transformation[7]. AI technology, through DL algorithms, has significantly improved the detection rate of early esophageal cancer, enabling early intervention and treatment[8]. In the case of gastroesophageal reflux disease (GERD), AI has been used to not only assist in diagnosis but also provide references for disease severity assessment and treatment selection by analyzing the degree of inflammation and extent of lesions in endoscopic images[9]. Additionally, AI can quickly detect esophagogastric varices and assess their risk of rupture, providing crucial information for clinical decision-making[10]. Recent studies have shown that the application of AI technology in the assessment of esophagogastric varices has resulted in significant progress. AI models based on DL can automatically identify the presence of esophagogastric varices by analyzing endoscopic images or videos. The detection accuracy of these models has reached 97.5% in multicenter studies, with a specificity of 97.8%, demonstrating extremely high diagnostic efficacy[10]. In addition, AI can construct predictive models for esophagogastric varices by integrating multiple clinical indicators. For instance, a non-invasive model based on computed tomography (CT) portography could predict high-risk varices that require treatment by analyzing parameters such as splenic volume and the diameter of the left gastric vein, with an area under the receiver operating characteristic curve of up to 0.799[11]. These technologies have not only improved the accuracy and efficiency of detection but also reduced the need for invasive conventional endoscopic examinations and the associated patient discomfort. Additionally, AI can support clinical decision-making by comprehensively analyzing factors such as the patient’s liver-function status, variceal diameter, and blood pressure to provide precise risk assessments and treatment recommendations[12].

Despite the immense potential of AI in the field of digestive endoscopy, many physicians remain skeptical about its accuracy and reliability, especially in diagnostic and treatment decision-making, where they tend to rely more on their own experience and judgment. This cautious attitude towards AI partly stems from the insufficient validation of existing technologies in real-world clinical environments and the lack of a fully defined collaborative model between AI and physicians. Therefore, to further promote the application of AI in digestive endoscopy, this paper reviews the current status and progress of AI in upper gastrointestinal diseases, focusing on its specific applications in BE, esophageal cancer, GERD, with the aim of providing references for clinical practice and research.

AI-ASSISTED DIAGNOSIS OF BE

BE refers to the pathological phenomenon where the squamous epithelium in the lower esophagus is replaced by columnar epithelium. Its diagnosis mainly relies on endoscopic examination and pathological biopsy. In recent years, AI technology has shown great potential in assisting the endoscopic diagnosis of BE. The development and potential of AI are generally realized through training, as shown in Figure 2.

Figure 2 The training process of artificial intelligence-assisted endoscopy. AI: Artificial intelligence.

AI not only helps endoscopists perform targeted biopsies for BE but also significantly improves the detection rate of BE[5]. Lee et al[6] used a convolutional neural network to identify early dysplasia in BE, and found that the network achieved a sensitivity of 91%, a specificity of 79%, and an area under the curve of 93% in detecting BE-related dysplasia in the test set. Additionally, the sensitivity for each lesion was 97% of the targeted biopsy results were consistent with expert and pathological evaluations. This indicates that AI can effectively identify early dysplasia in BE and assist endoscopists in performing precisely targeted biopsies, thereby reducing the risk of missed diagnoses and misdiagnoses.

Pan et al[7] further validated the potential of AI in BE diagnosis. By using 443 endoscopic images from 187 BE patients, the authors developed an AI model capable of automatically identifying and segmenting the extent of BE lesions. The model automatically annotated the gastroesophageal junction and the squamocolumnar junction in BE, with experimental results showing intersection over union (IoU) values of 0.56 and 0.82, respectively, achieving satisfactory results. This study demonstrated that AI can not only assist in diagnosis but also precisely annotate lesion areas in endoscopic images, providing more intuitive references for endoscopists. Struyvenberg et al[8] further expanded the application of AI in endoscopic diagnosis. Their computer-aided diagnosis (CAD) system analyzed 30021 individual endoscopic video frames. The video-based CAD system achieved an accuracy, sensitivity, and specificity of 83%, 85%, and 83%, respectively, in detecting BE, with an average evaluation speed of 38 frames per second. Additionally, in narrow-band imaging mode, the accuracy, sensitivity, and specificity of the CAD system for detecting BE were 84%, 88%, and 78%, respectively. These results indicate that AI systems can quickly process large volumes of endoscopic images while maintaining high diagnostic accuracy across different imaging modes. Abdelrahim et al[9] also validated the efficiency of AI in BE diagnosis. They used 7518 endoscopic images and videos of neoplastic BE from 96 patients, along with 1014973 endoscopic images and videos of non-neoplastic BE from 65 patients, as the training set to develop a CAD system for BE. Their system achieved a sensitivity, specificity, negative predictive value, and accuracy of 93.8%, 90.7%, 95.1%, and 92.0%, respectively, which were all significantly higher than the corresponding performance metrics of endoscopists (63.5%, 77.9%, 74.2%, and 71.8%). Moreover, their research was a multicenter study that validated the use of AI for real-time endoscopic video detection, further proving the potential of AI for widespread clinical application.

In a systematic review and meta-analysis[10], researchers analyzed 12 studies involving 1361 patients and 532328 endoscopic images. Their results showed that AI has high accuracy in detecting early neoplasia in BE, but they concluded that further research is needed on the role of AI in detecting the pathological subtypes of early neoplastic lesions in BE. Additionally, they considered that whether AI-assisted diagnosis is effective for both expert and non-expert endoscopists requires further investigation. Nevertheless, the above studies collectively indicate that AI has shown excellent performance in assisting BE diagnosis, and has great potential for clinical application.

AI IN THE DIAGNOSIS OF ESOPHAGEAL CANCER

Esophageal cancer is the seventh most common malignant tumor globally and the sixth leading cause of cancer-related deaths[11]. Squamous cell carcinoma and adenocarcinoma are the two most common pathological types of esophageal cancer, with squamous cell carcinoma being more prevalent in China[11]. Increasing endoscopic screening for high-risk populations and improving the diagnosis of early esophageal cancer can help reduce esophageal cancer-related mortality.

A study[12] developed a CAD system for the endoscopic detection of superficial esophageal squamous cell carcinoma. The CAD system achieved an accuracy of 92.9%, which was comparable to that of expert endoscopists (91.0%) and significantly higher than that of non-expert physicians (78.3%). When non-expert physicians referred to the AI results, their accuracy improved to 88.2%, demonstrating the advantage of AI in assisting endoscopists in the early diagnosis of esophageal cancer.

In a systematic review and meta-analysis[13], 28 studies involving 703006 endoscopic images of esophageal cancer were included. The results showed that DL models achieved an accuracy, sensitivity, specificity, positive predictive value, and negative predictive value of 92.90%, 93.80%, 91.73%, 93.62%, and 91.97%, respectively, in diagnosing esophageal cancer. Additionally, the diagnostic performance of the DL models in narrow-band imaging mode was higher than that in white-light mode. This suggests that DL models have great potential in accurately and rapidly diagnosing early esophageal cancer.

Feng et al[14] developed a new AI-assisted diagnostic system for superficial esophageal squamous cell carcinoma, which demonstrated excellent diagnostic performance. The system achieved a sensitivity of 90.17%, specificity of 94.34%, accuracy of 88.83%, positive predictive value of 89.5%, and negative predictive value of 94.72% in single-image analysis. In the human–computer comparison, the system showed a diagnostic performance comparable to that of expert endoscopists and significantly improved the accuracy, specificity, and positive predictive value of manual diagnosis. However, this study was a retrospective analysis of static images, and in the future, prospective studies based on dynamic videos will be needed to further validate the clinical utility of this AI-assisted system. Yuan et al[15] conducted a large-scale, multicenter, prospective, double-blind, randomized controlled trial in 12 hospitals in China. The study was based on 38409 white-light and non-magnified narrow-band images of superficial esophageal cancer and precancerous lesions, and developed an AI-assisted endoscopic system. The system significantly improved the detection rate of superficial esophageal cancer and precancerous lesions by endoscopists and reduced the rate of missed diagnoses.

AI IN THE DIAGNOSIS OF GERD

GERD is a common and complex clinical condition characterized by symptoms such as acid reflux, heartburn, and regurgitation. Based on its characteristic endoscopic and pathological features, GERD can be classified into reflux esophagitis, non-erosive reflux disease, and BE. Therefore, digestive endoscopy plays a crucial role in the diagnosis of GERD and the assessment of its severity[16]. Research on AI-assisted digestive endoscopy for GERD diagnosis is still limited, and mainly focuses on classification and segmentation.

One study on AI in GERD proposed a hierarchical heterogeneous descriptor fusion support vector machine framework to automatically identify GERD endoscopically, overcoming dimensionality issues[17]. However, this study did not perform endoscopic Los Angeles (LA) grading of esophagitis, and conventional machine learning methods are insufficient for analyzing large-scale, multi-category, high-resolution medical images. Another study[18] used support vector machine technology to classify lesions and non-lesions in esophagitis images, and assist in delineating the boundaries of esophagitis. However, the above study was based on only 227 endoscopic images of LA A-grade esophagitis lesions, and the boundaries of esophagitis lesions of LA grades C and D are more complex and difficult to distinguish by using AI, making the algorithm inadequate for clinical needs[18]. In 2021, Wang et al[19] proposed a DL model (GERD-vggnet) using convolutional neural networks to classify LA grades in white-light and narrow-band images of the gastroesophageal junction in GERD. The model achieved an accuracy of 87.9%. In 2022, the team combined DL and machine learning to develop the Gerd Net-RF model, which further improved the accuracy to 92.5% ± 2.1%[20]. In 2023, a study proposed an uncertainty-aware network that achieved 90.2% accuracy in classifying imbalanced esophagitis categories[21]. In the same year, another study proposed the Efficient Net model, which achieved a relatively better classification accuracy of 95.7%[22]. These studies were all based on LA grading, and classified images into three categories: LA grades A–B, LA grades C–D, and healthy esophagus. In clinical practice, the challenge of LA grading in GERD lies in precise classification between the A and B grades and between the C and D grades, making accurate LA grading essential.

Also in 2023, researchers for the first time used explainable AI to develop a five-grade classification DL model, Dense Net-121, based on LA grading. The model achieved an accuracy of 86.7%, which was significantly higher than that of senior endoscopists[23]. Other authors[24] have applied U-Net and its variants to the segmentation of endoscopic images of GERD, with the Attention U-Net model achieving an IoU value of 0.9007, showing promising results. However, the above study had some limitations, as the dataset consisted of only 688 Low-resolution endoscopic images that varied in size and quality, and did not cover all severity levels of GERD[24].

Another study[25] reported a dual-path framework (DCS-U-Net) combining segmentation and classification. The framework first used U-Net for coarse segmentation, then refined the segmentation results with a classifier. The experimental results showed that the precision of this framework was better than that of U-Net alone or a handcrafted feature-based approach. However, this study also had an incomplete dataset, containing only 795 endoscopic images of mild esophagitis and lacking clinically challenging and difficult-to-analyze images of LA grades C and D, making the DCS-U-Net framework insufficient for clinical needs[25]. Additionally, some literature suggests that the functional status of the gastroesophageal flap valve (GEFV) is closely related to GERD[26-28], indicating that GEFV grading may serve as supportive evidence for GERD diagnosis. Therefore, based on the Hill grading system for GEFV, a study proposed a four-class DL model, Res Net-50, combined with an attention mechanism; the model achieved a classification accuracy of 93.39%, demonstrating excellent Hill classification performance[29]. This is clinically meaningful for the diagnosis and treatment of GERD.

AI IN ESOPHAGOGASTRIC VARICES

Future research should include more multicenter, randomized trials and large-scale prospective studies to further validate the clinical application value of AI systems.

Esophagogastric variceal bleeding is one of the most fatal complications of liver cirrhosis and the most common cause of upper gastrointestinal bleeding. It is estimated that about 50% of liver cirrhosis patients develop esophagogastric varices, and about 30% of them experience variceal bleeding, with a mortality rate of 20%-30% for the first bleeding episode[30]. Endoscopy plays a crucial role in the diagnosis and treatment of esophagogastric varices, allowing direct observation of the morphology, size, and color of varices, as well as performing endoscopic ligation and sclerotherapy[31]. However, traditional endoscopy has limitations such as subjectivity and high operator dependency, requiring significant operator experience. In recent years, AI technology has shown great potential in assisting endoscopy for the diagnosis and treatment of esophagogastric varices, providing new approaches to address these issues.

In 2021, a study used 8566 endoscopic images of esophagogastric varices and 6152 normal esophagogastric images for training, and independent endoscopic images and videos for testing, to develop an AI system for real-time diagnosis of esophagogastric varices and prediction of bleeding risk. The system's accuracy in detecting esophageal and gastric varices was significantly better than that of endoscopists, and it could assist in identifying the size, morphology, and color of varices, as well as predicting endoscopic risk factors for variceal bleeding[32]. In 2022, the team further optimized the model, using 6034 endoscopic images of esophagogastric varices as the training dataset and validating it on 141 endoscopic videos and 11009 endoscopic images. Additionally, the study included endoscopic data from 161 liver cirrhosis patients for prospective validation, developing an XAI-based diagnostic system. The system outperformed endoscopists in diagnosing esophageal varices and was comparable to endoscopists in diagnosing gastric varices. The AI-assisted system also helped more patients receive appropriate clinical interventions, demonstrating its superiority in endoscopic-assisted diagnosis and treatment of esophagogastric varices[33]. Furthermore, radiomics and DL technologies have shown great potential in the non-invasive assessment of esophagogastric varices. For example, a DL model based on CT imaging can predict the severity and bleeding risk of esophagogastric varices through quantitative analysis. These technologies not only improve diagnostic accuracy but also provide non-invasive monitoring methods for clinical practice[34]. However, the application of AI in esophagogastric varices is still in its developmental stage, and current studies are mostly single-center or retrospective, with certain limitations.

In 2022, the team further optimized the model, using 6034 endoscopic images of esophagogastric varices as the training dataset and validating it on 141 endoscopic videos and 11009 endoscopic images[33]. Additionally, the study included endoscopic data from 161 liver cirrhosis patients for prospective validation, developing an XAI-based diagnostic system. The system outperformed endoscopists in diagnosing esophageal varices, with an accuracy of 93.8%, and was comparable to endoscopists in diagnosing gastric varices (90.8%). The AI-assisted system also helped more patients receive appropriate clinical interventions, demonstrating its superiority in endoscopic-assisted diagnosis and treatment of esophagogastric varices. The study also found that combining the AI system with endoscopists could further improve diagnostic accuracy, providing important evidence for the future development of human-computer collaborative diagnostic models.

Furthermore, radiomics and DL technologies have shown great potential in the non-invasive assessment of esophagogastric varices. For example, a DL model based on CT imaging can predict the severity and bleeding risk of esophagogastric varices through quantitative analysis[34]. The study included 419 liver cirrhosis patients and established a model to predict high-risk varices by measuring the spleen volume-to-platelet ratio (SVPR). The results showed that when SVPR > 3.78, the sensitivity and specificity for predicting high-risk varices were 80% and 74.4%, respectively. These technologies not only improve diagnostic accuracy but also provide non-invasive monitoring methods, helping to reduce unnecessary endoscopic examinations.

CONCLUSION

This review highlights the application progress of AI in the diagnosis of upper gastrointestinal diseases, emphasizing its potential in improving the accuracy of endoscopic image analysis, early lesion detection, and disease grading (Table 1). With the rapid development of AI technology, its application in gastroenterology has gradually transitioned from theoretical research to clinical practice, bringing unprecedented opportunities for the diagnosis, treatment, and prognosis assessment of gastric cancer and other upper gastrointestinal diseases. AI's application in endoscopic image analysis, pathological diagnosis assistance, radiological diagnosis, and treatment planning has significantly improved the detection rate and diagnostic accuracy of early lesions, providing strong support for personalized medicine. However, we must also recognize that the application of AI technology in gastroenterology is still in its exploratory stage, and its clinical promotion faces challenges such as data quality, insufficient multicenter validation, lack of real-time performance, and medical ethics. On one hand, the development and validation of AI models require high-quality, diverse datasets to ensure their generalization ability in different clinical scenarios. On the other hand, the clinical application of AI systems must undergo rigorous prospective, multicenter research validation to ensure their safety and effectiveness. Additionally, the interpretability and transparency of AI technology are key concerns for clinicians. Only when the decision-making process of AI can be understood and trusted by physicians can its clinical application truly take root. In future gastroenterology practice, AI is expected to become a valuable assistant to physicians, but not a replacement. The ultimate goal of AI technology is to provide clinicians with more accurate diagnostic tools, optimized treatment plans, and reliable prognosis assessments, thereby improving patient outcomes and quality of life. We look forward to further technological advancements and in-depth clinical research, enabling AI to play a greater role in the comprehensive management of gastrointestinal diseases and bringing new breakthroughs and developments to the field of gastroenterology.

Table 1 The summary table of the application progress of artificial intelligence in the diagnosis of upper gastrointestinal diseases.

Disease type	Application scenario	Research progress	Key indicators	Notes
BE	Assisted diagnosis, targeted biopsies	Convolutional Neural Network for detecting dysplasia in BE: Sensitivity 91%, specificity 79%, AUC value 93%[6]	Sensitivity: 91%-93.8%; specificity: 79%-90.7%; accuracy: 92.0%	AI significantly improves the detection rate of BE, reducing missed and misdiagnosed cases
		AI model for automatic identification and segmentation of BE extent: IoU values 0.56 and 0.82[7]
		Multicenter real-time endoscopic video detection: Sensitivity 93.8%, specificity 90.7%, accuracy 92.0%[9]
Esophageal cancer	Early lesion detection	CAD system for detecting squamous cell carcinoma: Accuracy 92.9%, comparable to experts[12]	Sensitivity: 93.8%; specificity: 91.73%; accuracy: 92.9%-94%	AI shows excellent performance in early esophageal cancer detection
		DL models: Accuracy 92.9%-94% in NBI mode, 92.9% in white-light mode[13]
		AI-assisted system improves detection rates and reduces missed diagnoses[15]
GERD	Disease classification, lesion segmentation	DL model for LA grading: Accuracy 87.9%-95.7%[19,22]	Accuracy: 87.9%-95.7%; IoU value: 0.9007	AI performs well in GERD grading and lesion segmentation
		XAI model for five-class classification: Accuracy 86.7%, higher than senior endoscopists[23]
		Attention U-Net segmentation model: IoU value 09007[24]
Esophagogastric varices	Real-time diagnosis, bleeding risk prediction	AI system for detecting esophagogastric varices: Accuracy 93.8%, comparable to endoscopists[33]	Accuracy: 93.8%; sensitivity: 80% (bleeding risk prediction); specificity: 74.4%	AI outperforms traditional methods in variceal diagnosis and risk prediction
		DL model based on CT imaging for predicting variceal bleeding risk[34]

BE: Barrett's esophagus; AI: Artificial intelligence; DL: Deep learning; IoU: Intersection over union; AUC: Area under the curve; CAD: Computer-aided diagnosis; CT: Computed tomography; GERD: Gastroesophageal reflux disease.